Firm Heterogeneity in Consumption Baskets: Evidence from Home and Store Scanner Data - The Review of ...

Page created by Joshua Rojas
 
CONTINUE READING
Firm Heterogeneity in Consumption Baskets:
                   Evidence from Home and Store Scanner Data∗
                                   Benjamin Faber† and Thibault Fally‡

                                                     May 2021

                                                     Abstract
       A growing literature has documented the role of firm heterogeneity within sectors for nominal
       income inequality. This paper explores the implications for household price indices across the
       income distribution. Using detailed matched US home and store scanner microdata, we present
       evidence that rich and poor households source their consumption differently across the firm
       size distribution within disaggregated product groups. We use the data to examine alternative
       explanations, propose a tractable quantitative model with two-sided heterogeneity that ratio-
       nalizes the observed moments, and calibrate it to explore general equilibrium counterfactuals.
       We find that larger, more productive firms sort into catering to the taste of richer households,
       and that this gives rise to asymmetric effects on household price indices. We quantify these
       effects in the context of policy counterfactuals that affect the distribution of disposable incomes
       on the demand side or profits across firms on the supply side.

       Keywords: Firm heterogeneity, real income inequality, household price indices, scanner data
       JEL Classification: F15, F61, E31

   ∗
      We would like to thank Mary Amiti, Costas Arkolakis, David Atkin, Kirill Borusyak, Pablo Fajgelbaum, Cecile
Gaubert, Jessie Handbury, James Harrigan, Xavier Jaravel, Thierry Mayer, Volker Nocke, Steve Redding, Andres
Rodriguez-Clare, Nicolas Schultz, Eric Verhoogen, Jonathan Vogel and participants at multiple seminars and con-
ferences for helpful comments and discussions. Dmitri Koustas and May-Lyn Cheah provided outstanding research
assistance. Researcher(s) own analyses calculated (or derived) based in part on data from The Nielsen Company
(US), LLC and marketing databases provided through the Nielsen Datasets at the Kilts Center for Marketing Data
Center at The University of Chicago Booth School of Business. The conclusions drawn from the Nielsen data are
those of the researcher(s) and do not reflect the views of Nielsen. Nielsen is not responsible for, had no role in, and
was not involved in analyzing and preparing the results reported herein.
    †
      Department of Economics, UC Berkeley and NBER.
    ‡
      Department of Agricultural and Resource Economics, UC Berkeley and NBER.
1         Introduction
Income inequality has been on the rise in the US and many other countries, attracting the sus-
tained attention of policy makers and the general public. In this context, a growing literature has
documented the role of Melitz-type firm heterogeneity within sectors in accounting for nominal
income inequality.1 In this paper, we explore the implications of firm heterogeneity for household
price indices across the income distribution. We aim to answer three central questions: i) to what
extent do rich and poor households source their consumption baskets from different parts of the
firm size distribution?; ii) what forces explain these differences?; and iii) what are the implications
of the answers to i) and ii) for the impact of policies or economic shocks on real income inequality?
    In answering these questions, the paper makes three main contributions. First, using detailed
matched home and store scanner consumption microdata, we document significant differences
in the weighted average firm sizes that rich and poor US households source their consumption
from, and explore alternative explanations.2 Second, to rationalize these moments we develop a
tractable quantitative model of product quality choice with two-sided heterogeneity across firms
and households, and use the microdata to calibrate it and quantify the forces underlying the
stylized facts. Third, we explore general equilibrium (GE) policy counterfactuals to illustrate
how, in a setting where rich and poor households source their consumption from heterogeneous
firms, policies and economic shocks give rise to asymmetric price index effects across the income
distribution.
    At the center of the analysis lies a detailed collection of microdata that allow us to trace the
firm size distribution into the consumption baskets of households across the income distribution.
We combine a dataset of 345 million consumer transactions when aggregated to the household-
by-retailer-by-barcode-by-half-year level from the Nielsen US Home Scanner (Nielsen Company,
2016a) data over the period 2006-2014, with a dataset of 12.2 billion store transactions when
aggregated to the store-by-barcode-by-half-year level from the Nielsen US Retail Scanner data
(Nielsen Company, 2016b) covering the same period. The combination of home and store-level
scanner data allows us to trace the size distribution of producers of brands—measured in terms of
national sales that we aggregate across on average 27,000 retail establishments each half year in the
store scanner data—into the consumption baskets of on average 59,000 individual households per
half year in the home scanner data within more than 1000 disaggregated retail product modules,
such as carbonated drinks, shampoos, pain killers, desktop printers or microwaves.
    The analysis proceeds in four steps. In step 1, we use the data to document a set of stylized facts.
We find that the richest 20 (resp. 10) percent of US households source their consumption from
on average 20 (resp. 27) percent larger producers of brands within disaggregated product groups
compared to the poorest 20 (resp. 10) percent of US households. We also document that these
differences in weighted-average firm sizes arise in a setting where the rank order of household budget
shares on different producers within a product group is preserved across income groups—i.e. the
largest firms command the highest budget shares for all income groups. We interpret these stylized
facts as equilibrium outcomes in a setting where both consumers and firms optimally choose their
product attributes. We also use the microdata to explore a number of alternative explanations,
    1
        e.g. Card et al. (2013); Helpman et al. (2017); Song et al. (2018). See related literature at the end of this section.
    2
        We refer to firm size in terms of relative firm sales within product groups.

                                                                1
and find that differential coverage of retail consumption or differences in product supply and
pricing across the income distribution are unlikely driving these results. We also document that
the relationship between incomes and firm sizes holds across different product departments in the
data, and explore whether consumption categories not fully represented in the Nielsen data, such
as consumer durables, health expenditures or digital media, show similar patterns. Focusing on
subsets of these categories that are covered in the data—household appliances, pharmaceuticals and
video, audio and software purchases respectively—we find similar or more pronounced differences
in firm sizes across incomes.
    In step 2 we write down a tractable model that rationalizes the stylized facts in the data.
On the consumption side, we specify non-homothetic preferences allowing households across the
income distribution to differ both in terms of their price elasticities as well as in their evaluations
of product quality attributes. On the production side, we introduce quality choice into a model
of heterogeneous firms within sectors. Both marginal and fixed costs can be functions of output
quality, allowing for economies of scale in production. Markups can vary across firms due to
both oligopolistic competition and selling to heterogeneous consumers. Firms now operate in
a setting where their pricing and quality choices affect the composition of market demand that
they face. Modeling product choices with two-sided heterogeneity implies that shocks that affect
producers differently, such as trade integration, corporate taxes and regulations, can feed into
the consumption baskets of rich and poor households asymmetrically. In turn, changes in the
distribution of disposable incomes, due to e.g. income tax reform, affect firms differently across
the size distribution with GE knock-on effects on household price indices.
    In step 3, we use the microdata to estimate the preference and technology parameters. On the
demand side, we find that rich and poor households differ both in terms of price elasticities and
their valuation of product quality attributes. We find that poorer households have higher price
elasticities relative to richer households, but that these differences, while statistically significant,
are relatively minor. We also find that, while households on average agree on the ranking of
quality evaluations across producers, richer households value higher quality significantly more. On
the production side, we estimate that producing higher quality increases both the marginal and the
fixed costs of production, giving rise to economies of scale in quality production. To estimate the
technology parameters, we use two different estimation strategies. The first follows existing work,
and is based on cross-sectional variation in firm scale and output quality. The second exploits
within-firm changes in brand quality and scale over time. To identify the effect of firm scale on
output quality in the panel estimation, we use state-level measures of changes in brand quality on
the left-hand side, and construct an instrument (IV) for national brand scale on the right-hand
side. The IV exploits pre-existing differences in the geography of brand sales across other US states
interacted with state-level variation in average sales growth observed in other product groups.
    The parameter estimates from step 3 reveal two opposing forces that in equilibrium determine
both the sorting of firms across product quality attributes and firm sizes across consumption
baskets. On the one hand, larger firms offer lower quality-adjusted prices, which increases the share
of their sales coming from more price-sensitive lower-income consumers. Since these consumers
value quality relatively less, this channel, ceteris paribus, leads poorer households to source their
consumption from on average larger firms, which in turn pushes these firms to produce at lower

                                                   2
output quality. On the other hand, the estimated economies of scale in quality production give
larger firms incentives to sort into higher product quality, catering to the taste of richer households.
Empirically, we find that this second channel dominates the first, giving rise to the endogenous
sorting of larger, more productive firms into products that are valued relatively more by richer
households. Armed with these estimates, we find that the observed stylized facts from step 1
translate into significant differences in the weighted-average product quality and quality-adjusted
prices embodied in consumption baskets across the income distribution. The richest 20 percent
of US households source their consumption from on average 22 percent higher-quality producers
compared to the poorest quintile of households. At the same time, we find that the richest income
quintile source their consumption at on average 10 percent lower quality-adjusted prices. Overall,
we find that the calibrated model based on the estimates from step 3 fits the observed differences
in firm sizes across consumption baskets from step 1 both qualitatively and quantitatively.
    In the final step 4, we use the calibrated model to quantify a new set of price index implications
in the context of two policy counterfactuals. The first counterfactual affects the distribution of
disposable incomes on the demand side, and is motivated by the recent debate about progressive
income taxation in the US and elsewhere. We evaluate the price index effects of returning from
current US tax rates to a more progressive post-WWII US tax schedule, increasing the effective
rate on the richest household group in our calibration from currently around 30 to 50 percent.
This policy also closely corresponds to the counterfactual of moving the US to the current average
effective rate on this group among Northern European countries, and it is in line with the proposed
tax reforms of two presidential candidates for the 2020 US elections (Sanders and Warren). We find
that the resulting compression of disposable incomes gives rise to GE knock-on effects—through
changes in firm scale, output quality, variable markups and exit/entry across the firm size distribu-
tion—that affect rich and poor households differently, amplifying the progressivity of the reform.
As a result, the poorest 20 percent of US households experience a 3 percentage point lower inflation
rate for retail consumption compared the richest 20 percent.3
    The second policy counterfactual affects the profits of firms across the size distribution on
the supply side. The counterfactual is motivated by the ongoing debate about closing loopholes
in corporate taxation. A growing literature in public finance has documented larger possibilities
for tax avoidance or evasion among large US corporations (e.g. Bao & Romeo (2013), Guvenen
et al. (2017), Wright & Zucman (2018)). We use these findings to evaluate the implications of
eliminating the kink that has been documented at the 95th percentile of the firm size distribution
in the otherwise smoothly increasing relationship between firm size and effective corporate tax
rates. This policy would lead to an increase of on average 5 percent in corporate taxes paid by the
largest 5 percent of producers, ranging between on average 1 percent at the 95th percentile to 11
percent at the 99th. We find that even such a relatively modest adjustment in corporate taxation
leads to a meaningful GE effect on inflation differences between rich and poor households. This is
in the order of a 1.5 percentage point lower cost of living increase for retail consumption among
the bottom 20 percent of US households compared to the top 20 percent. We document that the
direct incidence of this policy—holding initial household and firm decisions fixed—accounts for
   3
     The data allow us to calibrate the model at the level of five broad income groups, while both tax changes and
taste for quality increase with incomes within these groups. Results are conservative in this respect.

                                                        3
about one third of the inflation difference, with the remainder due to endogenous firm adjustments
that affect consumption baskets differently.
     We also explore the implications for the distribution of the gains from trade, extending a text-
book Melitz model with two symmetric countries to quality choice with two-sided heterogeneity
within countries, and calibrating it to the US microdata. We find that allowing for firm hetero-
geneity across consumption baskets makes the gains from trade more unequal: increasing bilateral
import penetration by 10 percentage points leads to a 3.5 percentage point lower retail inflation
for the richest 20 percent of households compared to the poorest 20 percent. Taken together, these
findings illustrate that firm heterogeneity affects inequality in more complex ways than through
the nominal earnings of workers that have been the focus of the existing literature. These insights
arise after introducing a basic set of features that we observe in the microdata—allowing for prod-
uct attribute choice by heterogeneous firms and households—into an otherwise standard economic
environment.
     This paper is related to the growing literature on the extent, causes and consequences of firm
heterogeneity within sectors that has spanned different fields in economics, including international
trade (Bernard et al., 2007; Melitz, 2003), industrial organization (Bartelsman et al., 2013), macroe-
conomics (Hsieh & Klenow, 2009), development (Peters, 2013), labor economics (Card et al., 2013)
and management (Bloom & Van Reenen, 2007). Within this literature, this paper is most closely
related to existing work on the implications of firm heterogeneity for nominal income inequality
(Burstein & Vogel, 2017; Card et al., 2013; Davis & Harrigan, 2011; Frias et al., 2009; Helpman et
al., 2017; Sampson, 2014; Song et al., 2018).
     Our theoretical framework builds on existing work, both on quality choice across the domestic
income distribution with non-homothetic preferences on the demand side (Fajgelbaum et al., 2011;
Handbury, 2019) and on quality choice across heterogeneous firms on the supply side (e.g. Bastos
et al. (2018); Feenstra & Romalis (2014); Fieler et al. (2018); Johnson (2012); Kugler & Verhoogen
(2012)). While these building blocks have been used separately—i.e. heterogeneous firms facing a
representative agent in each country or non-homothetic preferences facing homogeneous firms—this
paper combines them in a quantitative GE model. This two-sided heterogeneity within countries
allows us to rationalize the observed differences in firm sizes across consumption baskets, and gives
rise to new price index implications across the income distribution in the light of economic shocks
or policy changes.4
     The paper is also related to a growing empirical literature using the Nielsen data (Broda &
Weinstein, 2010; Handbury, 2019; Handbury & Weinstein, 2014; Hottman et al., 2016). Most of
this literature has been based on the home scanner data. More recently, Argente & Lee (2016) and
Jaravel (2019) use the home and store scanner data to document that lower-income households
have experienced higher inflation over the past decade and beyond. Argente & Lee (2016) relate
   4
    A notable exception is Feenstra & Romalis (2014) who combine quality choice by heterogeneous firms under
monopolistic competition with non-homothetic preferences on the demand side. Since their objective is to infer
quality from observed unit values in international trade flows, their model features a representative agent within
countries on the demand side. On the supply side, they follow e.g. Baldwin & Harrigan (2011); Kugler & Verhoogen
(2012) modeling quality choice as a deterministic function of a firm’s productivity draw. To be able to rationalize
our empirical results on the effect of firm scale on quality upgrading in firm-level panel data, we instead allow fixed
costs to be a function of output quality, following earlier work by Sutton (1998), Hallak & Sivadasan (2013) and the
second model variant in Kugler & Verhoogen (2012).

                                                          4
this finding to a higher possibility for quality-downgrading among higher-income households during
the Great Recession, and Jaravel (2019) to more innovation and competition in product groups
consumed relatively more by richer households. In this paper, we present new empirical evidence
suggesting that the widely documented presence of firm heterogeneity within sectors translates
asymmetrically into the consumption baskets of rich and poor households, quantify the underlying
forces and explore a new set of implications in the context of policy counterfactuals.
    The remainder proceeds as follows. Section 2 describes the data. Section 3 documents stylized
facts and explores alternative explanations. Section 4 presents the model. Section 5 presents
parameter estimation. Section 6 presents the counterfactual analysis. Section 7 concludes.

2        Data
Retail Scanner Data The Nielsen retail scanner data consist of weekly price and quantity
information generated by point-of-sale systems for more than 100 participating retail chains across
all US markets between January 2006 and December 2014. When a retail chain agrees to share
their data, all of their stores enter the database. As a result, the database includes more than
50,000 retail establishments. The stores in the database vary widely in terms of formats and
product types: e.g. grocery, drug, mass merchandising, appliances, liquor, or convenience stores.
    Data entries can be linked to a store identifier and a chain identifier so a given store can be
tracked over time and can be linked to a specific chain. While each chain has a unique identifier, no
information is provided that directly links the chain identifier to the name of the chain. This also
holds for the home scanner dataset described below. The implication of this is that the product
descriptions and barcodes for generic store brands within product modules have been anonymized.
However, both numeric barcode and brand identifiers are still uniquely identified, which allows us
to observe sales for individual barcodes of generic store brands within each product module in the
same way we observe sales for non-generic products.
    When aggregated to the store-by-barcode-by-half-year level, each half year covers on average
$113 billion worth of retail sales across 27,000 individual stores in more than 1000 disaggregated
product modules, 2500 US counties and across more than 730,000 barcodes belonging to 175,000
producers of brands (see Appendix Table A.1).5 As described in more detail in the following
section, we use these data in combination with the home scanner data described below in order to
trace the distribution of firm sizes (measured in terms of national sales measured across on average
27,000 stores per half year) into the consumption baskets of individual households.

Home Scanner Data The Nielsen home scanner data is collected through hand-held scanner
devices that households use at home after their shopping in order to scan each individual transac-
tion they have made. Importantly, the home and store level scanner datasets can be linked: they
use the same codes to identify retailers, product modules, product brands as well as barcodes. As
described in more detail in the following section, we use this feature of the database to estimate
weighted average differences in firm sizes across consumption baskets.
    5
        We do not make use of Nielsen’s “Magnet” database that covers non-barcoded products, such as fresh produce.

                                                          5
When aggregated to the household-by-barcode-by-half-year level, each half year covers on av-
erage $109 million worth of retail sales across 59,000 individual households in more than 1000
product modules, 2600 US counties and close to 600,000 barcodes belonging to 185,000 producers
of brands (Appendix Table A.1). One shortcoming of the home scanner dataset is that nominal
household incomes are measured imprecisely. First, incomes are reported only across discrete in-
come ranges. And those income bins are measured with a two-year lag relative to the observed
shopping transactions in the dataset. To address this issue, we divide households in any given
half year into percentiles of total retail expenditure per capita.6 To address potential concerns
about a non-monotonic (or decreasing) relationship between total retail outlays and incomes, we
also confirm that our measure of total retail expenditure per capita is monotonically increasing
in reported nominal incomes two years prior (confirming existing evidence that retail expenditure
has a positive income elasticity) (Appendix Figure A.1).
    These descriptive statistics also help clarify the relative strengths and weaknesses of the two
Nielsen datasets. The strength of the home scanner database is the detailed level of budget
share information that it provides alongside household characteristics. Its relative weakness in
comparison to the store-level retail scanner data is that the home scanner sample of households
only covers a small fraction of the US retail market in any given period. Relative to the home
scanner data, the store-level retail scanner data cover more than 1000 times the retail sales in each
half year. This paper takes advantage of both datasets for the empirical analysis, by combining
national sales by product from the store scanner data with the detailed information on individual
household consumption shares in the home scanner data.

3       Stylized Facts
This section draws on the combination of the home scanner and retail scanner data to document a
set of stylized facts about firm heterogeneity in the consumption baskets of households across the
income distribution. Online Appendix figures and tables are prefixed by “A.”.

3.1     Definition of Firms
Using both datasets, we can document what has been shown many times in manufacturing estab-
lishment microdata (Bartelsman et al., 2013; Bernard et al., 2007): firm sizes (measured here in
terms of sales) differ substantially within disaggregated product groups (Figure A.2). We define
a firm as a producer of a brand within product modules in the Nielsen data. This leads to an
average of about 150 active firms within a given product module. Two possible alternatives would
be to define a firm as a barcode product (leading to on average 700 firms per module), or as a
holding company (leading to on average less than 40 firms per module). To fix ideas, an example
for the product module Shampoo would be the barcode product Ultimate Hydration Shampoo (22
    6
     Per capita expenditure can be misleading due to non-linearities in per capita outlays with respect to household
size (e.g. Subramanian & Deaton (1996)). To address this concern, we non-parametrically adjust for household size
by first regressing log total expenditure on dummies for each household size with a household size of 1 being the
reference category and a full set of household socio-economic controls. We then deflate observed household total
expenditure to per capita equivalent expenditure by subtracting the point estimate of the household size dummy
(which is non-zero and positive for all households with more than one member).

                                                         6
oz bottle) belonging to brand TRESemmé that is owned by the holding company Unilever.
     We choose the definition of firms as brands within product modules for two main reasons, and
then check the robustness of our findings to alternative definitions. First, our objective is to define
a producer within a given module as closely as possible to an establishment in commonly used
manufacturing microdata, following the bulk of the recent literature on firm heterogeneity (Melitz,
2003). The definition of firms as holding companies (e.g. Procter&Gamble) would be problematic
as these conglomerates operate across hundreds of brands produced in different establishments. The
definition of firms at the barcode level would be problematic for the opposite reason, because the
same establishment produces different pack sizes of the same product that are marked by different
barcodes. In this light, defining producers of brands within disaggregated product modules as firms
is likely the closest equivalent to observing several different establishments operating in the same
disaggregated product group. Second, our theoretical framework features endogenous investments
in product quality across firms, and it is at the level of brands within product groups that these
decisions appear to be most plausible.7
     The distribution of national market shares measured using the home scanner data (on average
59,000 households per half year) appears to be compressed relative to that measured using the store
scanner data (on average 27,000 thousand stores per half year) (Figure A.2). This compression
is stronger before applying the Nielsen household sampling weights, but still clearly visible after
applying the weights. It also holds when restricting attention to producers of brands observed
in both datasets. In the following, we report the main new stylized fact using the firm size
distributions computed from both datasets. Given that the retail scanner data capture more than
1000 times the amount of transactions compared to the home scanner data, we then choose the
store scanner data as our preferred measure of brand-level national market shares.

3.2    Firm Heterogeneity Across Consumption Baskets
Figure 1 depicts the main stylized fact of the paper. Pooling repeated cross-sections across 18
half-year periods, we depict percentiles of household per capita expenditure on the x-axis and
weighted average deviations of log firm sales from the product module-by-half-year means on
the y-axis.8 The weights correspond to each household’s retail consumption shares across all
brands in all product modules consumed during the six-month period. When collapsed to five per
capita expenditure quintiles on the right panel of Figure 1, we find that the richest 20 percent of
US households source their consumption from on average 20 percent larger producers of brands
within disaggregated product modules compared to the poorest 20 percent. These figures are our
preferred measure of the national firm size distribution using the store scanner data. But, as
the figure shows, a very similar relationship holds when using the firm size distribution from the
home scanner data instead. This relationship is monotonic across the income distribution, and the
   7
     To corroborate this, we confirm in the data that 95 percent of variation in the average unit value paid for
barcodes within product modules is accounted by for by brand-by-pack-size fixed effects. Similarly, on average 80
percent of the sum of absolute differences in expenditure shares between the richest and poorest household quintiles
across all barcodes within product modules are accounted for by differences in budget shares across brands (leaving
20 percent to be explained by differences within brands).
   8
     To avoid measurement error from exiting or entering households in the consumer panel, we restrict attention to
households for each half-year period that we observe to make purchases in both the first and the final month of the
half year.

                                                         7
firm size difference increases to 27 percent when comparing the richest and poorest 10 percent of
households. As discussed above, Figure 1 is also robust to alternative definitions of firms in the
Nielsen data: Figure A.3 shows close to identical results when defining firms instead as holding
companies within product groups.9
     What types of shopping decisions are driving these differences in weighted average firm sizes
across the income distribution? In Table A.2, we present the brands with the most positive and
most negative differences in consumption shares between rich and poor household quintiles across
three popular product modules for each of the eight product departments in our consumption
microdata. We also list the difference in their log average unit values (price per physical unit) as
well as the difference in their national market shares within that product module. Two features
stand out. First, the brand that is most disproportionately consumed by the rich has a higher unit
value and a larger market share relative to the brand that is most disproportionately consumed by
the poor.10 Second, looking at the brand names it seems that richer households have a tendency
to consume from the leading premium brands whereas the poorest quintile of households have a
higher tendency to pick either generic store brands, or cheaper second and third-tier brands in the
product group (e.g. Tropicana vs generic OJ, Pepsi vs generic Cola, Duracell vs Rayovac, Tide vs
Purex, Dove vs Dial, Heinz vs Hunt’s).
     Finally, we investigate whether these observed differences in product choices are driven by a
fundamental disagreement about relative product appeal across rich and poor households. Do
we see rich households consuming a large share of their expenditure from the largest producers
while poor households spend close to none of their budget on those same producers? Or do
households from different income groups agree on their relative evaluations of value-for-money
across producers, such that the rank order of their budget shares is preserved across the income
distribution? Figure A.6 documents that the latter appears to be the case in the data. The figure
depicts—separately for each income group—non-parametric estimates of the relationship between
income group-specific deviations in log expenditures across brands within product modules on the
y-axis and deviations in log total market sales of those same producers in the store scanner data on
the x-axis. The fact that expenditure shares within each income group increase monotonically with
total firm size suggests that households on average agree on their evaluation of product quality
attributes given prices, and that the rank order of budget shares across producers is preserved to a
striking extent across all income groups. To express this in a single statistic, we find that the rank
order correlation between the richest income quintile and the poorest for rankings of brand market
shares within product modules is .89 when pooled across all product modules in the data. However,
it is also apparent from the difference in slopes depicted in Figure A.6 that while all households
spend most of their budget on the largest firms within product modules, richer households spend
relatively more of their total budget on these largest producers relative to poorer households.
     Related to this, we also investigate the role of the extensive margin in product choice underlying
Figure 1. We find that while the average number of both UPCs and brands consumed within a
   9
     Figure A.4, we shows an alternative representation plotting sales shares from different quintiles of the income
distribution on the y-axis across firm sizes on the x-axis. Figure A.5 shows the relationship in Figure 1 after
computing firm sizes in terms of quantities (units sold).
  10
     The scanner data allow the comparison of prices per unit (unit values) for identically measured units across
products within product module (e.g. liters of milk, units of microwaves, grams of cereal, etc).

                                                         8
module increase with incomes, the fraction of total retail expenditure accounted for by products
consumed across all 5 income groups is very large (> 95 percent) and close to constant across
the income distribution (Figure A.7). Consistent with the rank order correlations above, this
suggests that the extensive margin, while visible in the data, is unlikely to play a significant role
in accounting for differences in average firm sizes across groups of rich and poor households.11

3.3    Interpretation and Alternative Explanations
One natural interpretation of these stylized facts is that they arise as equilibrium outcomes in a
setting where both heterogeneous firms and households choose the product attributes they produce
or consume. However, there are a number of alternative and somewhat more mechanical expla-
nations that we explore using the microdata before moving on to the model. Another question is
how representative these findings are for consumption categories that are not well represented in
the Nielsen data.

Representativeness and Data-Driven Explanations One concern is that the relationship
documented in Figure 1 could in part be driven by shortcomings of the data. First, we explore
the representativeness of the stylized fact in Figure 1. To this end, we estimate the relationship
separately for each of the 18 half-year periods and for each of eight broad product categories in the
Nielsen data: beverages, dairy products, dry grocery, frozen foods, general merchandise, non-food
grocery, health and beauty, and packaged meat (Figure A.9).12 We find that the pattern of firm
size differences across consumption baskets holds across these different product segments and is
not driven by one particular type of consumer products. We also find that the stylized fact in
Figure 1 holds in each of the 18 half-year periods in our data.
    As reported by e.g. Broda & Weinstein (2010) and more recently in Jaravel (2019), the product
groups covered in the Nielsen data can be matched to product groups in the US CEX expenditure
surveys that acount for about 40 percent of goods consumption, which in turn accounts for about
one third of total household outlays. To further explore to what extent the stylized fact in Figure
1 holds among consumption categories not fully represented in the Nielsen data, such as durables
consumption, health expenditures or digital media, we use parts of these categories that are covered
in the data. In particular, we estimate Figure 1 separately for household appliances (e.g. air
conditioners, refrigerators, kitchen appliances, desktop printers), pharmaceuticals purchases and
audio, video and software purchases (Figure A.10). Reassuringly, we find that the relationship
holds similarly or is more pronounced.13
    Second, it could be the case that generic store brands are produced by the same (large) pro-
ducers and sold under different labels across retail chains. If poorer households source more of
their consumption from generics, then we could under-estimate their weighted average producer
   11
      In support of this, Figure A.8 confirms that we find close to identical results after limiting the product space to
only products consumed by all income quintiles in a given period (or across all periods).
   12
      We combine observations for alcoholic and non-alcoholic beverages as one department in these graphs. Our
reported findings above hold separately for both of these departments. We pool them here to be consistent with
Section 5, where having one combined group for Beverages addresses sparsity in the parameter estimation.
   13
      The Nielsen data unfortunately do not cover services consumption. For future research, newly available credit
card microdata could potentially be used to investigate the relationship between card holder incomes and the firm
size of service providers they source their consumption from.

                                                           9
size due to this labeling issue. To address this concern, we re-estimate Figure 1 after restricting
consumption to sum to 100% for all non-generic product consumption for each household. We find
close to identical results compared to the baseline in Figure 1 (Figure A.11).
    Third, it could be the case that we are missing systematically different shares of retail con-
sumption across rich and poor households due to the exclusion of products sold by retail chains
that are not participating in the store-level retail scanner data (that we use to compute national
market sales across producers). For example, it could be the case that richer households purchase
a larger share of their retail consumption from independent boutique retailers selling small special-
ized brands, in which case we would be over-stating the weighted average firm sizes among richer
households. Conversely, it could be true that richer households spend on average more on retail
chains (compared to e.g. corner stores). To address such concerns, we make use of the fact that
the home scanner data do not restrict reporting to participating retail chains in the store scanner
data.14 We then re-estimate Figure 1 after only including households for which we observe more
than 90, 95 or 97.5 percent of their total reported retail expenditure on brands that are matched
across both data sets, and find that the relationship holds close to identically in all specifications
(Figure A.12). To further corroborate, we also show that the fraction of total expenditure on
brands in both datasets differs by less then 3 percent across the income groups (Figure A.13).15
    Another data-related concern is that the Nielsen data do not allow us to observe firm sales out-
side the US market. For both US-based exporters and imported brands, we are thus mismeasuring
total firm sales relative to domestic-only US producers. Given existing evidence on the selection of
firms into trade, it is likely that the resulting measurement error in firm sizes is positively related to
the observed US market shares in the Nielsen data (understating true differences in firm sizes). To
corroborate this, we report differences in weighted-average firm sizes across incomes separately for
product groups both below and above-median import penetration or export shares.16 We find that
the differences in firm sizes across rich and poor households are indeed slightly more pronounced
in the below-median sectors for both import or export shares (Figures A.14 and A.15).

Differences in Supply and Pricing Across Incomes Another explanation could be that rich
and poor households live in geographically segmented markets and/or shop across segmented store
formats, so that differential access to producers, rather than heterogeneous household preferences,
could be driving the results. We explore to what extent differences in household geographical
location as well as differences in retail formats within locations play in accounting for Figure 1
(Figure A.16). We first re-estimate the same relationship after conditioning on county-by-half-
year fixed effects when plotting the firm size deviations on the y-axis. Second, we additionally
condition on individual household consumption shares across 79 different retail store formats (e.g.
supermarkets, price clubs, convenience stores, pharmacies, liquor stores).17 We find a very similar
  14
      Since retail chain participation in the store scanner data is not made public by Nielsen, home scanner participants
are not made aware of this either.
   15
      The “missing retailers” concern is also not apparent in Figure 1 that depicts very similar patterns when using
100 percent of household retail consumption as reported in the home scanner data.
   16
      To this end, we match the Nielsen product groups to 4-digit SIC codes in 2005 US trade data. See Table A.3.
Below median is equivalent to less than 10 percent for both import penetration and export shares.
   17
      We condition on 79 store formats within the same county to capture potential differences in access across inner-
city vs. suburbs or due to car ownership. Conditioning on individual stores would give rise to the concern that
households choose to shop at different retailers precisely due to the product mix on offer, rather than capturing

                                                           10
relationship compared to Figure 1 in both cases, suggesting that differential access to producers is
unlikely to be the driver.
    Related to differential access to producers, differential pricing across income groups could be
another potential explanation. For this to matter in our context, it would have to be the case that
larger and smaller producers differ in their extent of price discrimination across income groups,
such that larger firms become relatively more attractive for richer households. To explore this in
the data, we plot the relationship between unit values and firm sizes separately for prices paid by
the poorest and the richest income quintiles (Figure A.17). If it were true that larger firms offer
differentially lower or higher prices compared to smaller firms across high and low-income groups,
we would expect the slope of these relationships to differ between the income groups. Using unit
values and firm sizes defined either at the UPC or at the brand-level, we find that this relationship
is close to identical for both rich and poor consumers, providing some reassurance that richer
households do not pick larger firms because of differential pricing.

Fixed Product Attributes Finally, we explore the notion that large firms are large because
they sell to richer households. If firms were born with fixed product attributes and/or brand
perceptions, and some firms got lucky to appeal to the rich, while other producers cannot respond
over time by altering their own product attributes or brand perceptions, this would mechanically
lead to richer households sourcing from larger firms (as the rich account for a larger share of total
sales).18
    Here, we document that in the medium or long run this notion seems hard to reconcile with
either the data or the existing literature on endogenous quality choice by firms. First, a body
of empirical work has documented that firms endogenously choose their product attributes as a
function of market demand in a variety of different empirical settings (e.g. Bastos et al. (2018),
Dingel (2016)).19 Second, the scanner data suggest that producers of brands frequently alter the
physical characteristics and/or presentation of their products over time. We find that during each
half year close to 10 percent of producers of brands replace their products with changed product
characteristics (e.g. packaging or product improvements) that have the identical pack sizes to the
previous replaced varieties on offer by the same brand (Table A.4)—suggesting that producers
are indeed capable of choosing their product attributes as a function of market conditions.20 In
support of these descriptive moments, we also provide more direct empirical evidence in Section
5 as part of our technology parameter estimation, documenting that an exogenous increase in the
differences in access.
   18
      This also relates to the original note in Melitz (2003) that the heterogeneity parameter can either be thought of
as a marginal cost draw in a setting with horizontal differentiation, or as a quality draw in a setting with vertical
differentiation.
   19
      Another literature in support of this is the marketing literature on firm strategies using advertising to affect
brand perceptions over time (e.g. Keller et al. (2011)).
   20
      It could still be the case that our 18 repeated cross-sections (half years) depicted in Figure 1 are partly capturing
the result of short-term taste shocks across products that differ between rich and poor households while hitting a
fixed number of producers with fixed product attributes. To further investigate this possibility, we re-estimate the
relationship in Figure 1 after replacing contemporary differences in firm sales by either sales three years before or
three years in the future of the current period. If the distribution of firm sizes was subject to significant temporary
swings over time, then we would expect the two counterfactual relationships to slope quite differently from our
baseline estimate in Figure 1. Instead, Figure A.18 suggests that the estimated differences in producer sizes are
practically identical.

                                                            11
scale of production leads to brand-level quality upgrading over time (see Tables 2 and 3).

4        Theoretical Framework
This section develops a tractable quantitative model that rationalizes the observed moments in the
microdata. We introduce two basic features into an otherwise standard Melitz model of hetero-
geneous firms. On the demand side, we allow for non-homothetic preferences so that consumers
across the income distribution can differ in both their price elasticities and in their product quality
evaluations. On the producer side, firms with different productivities face the observed distribution
of consumer preferences and optimally choose their product attributes and markups. Reflecting the
asymmetry of market shares within sectors we document above, we also depart from the assump-
tion of monopolistic competition, and allow for variable markups under oligopolistic competition.
The exposition below abstracts from time subscripts. Online Appendices 2-5 provide additional
details on the model and its extensions.

4.1      Model Setup
Consumption The economy consists of two broad sectors: retail consumption (goods available
in stores and supermarkets) and an outside sector. As in Handbury (2019), we consider a two-tier
utility where the upper-tier depends on utility from retail shopping UG and the consumption of an
outside good z:
                                         U = U (UG (z), z)                                    (1)

For the sake of exposition, we do not explicitly specify the allocation of expenditures in retail
vs. non-retail items, but assume that the outside good is normal.21 We denote by H(z) the
cumulative distribution of z across households in equilibrium, and normalize to one the population
of consumers. Utility from retail consumption is then defined by:

                                                                               αn (z).    σn (z)
                                                                                           σn (z)−1
                                        Y        X                   σn (z)−1
                             UG (z) =               (qni ϕni (z))    σn (z)                                      (2)
                                         n    i∈Gn

where n refers to a product module in the Nielsen data and i refers to a brand producer within
the product module.22 ϕni (z) is the perceived quality (product appeal) of brand i at income
level z, and σn (z) is the elasticity of substitution between brands in module n at income level z.
Allowing demand parameters for retail consumption to be functions of outside good consumption
z introduces non-homotheticities in a very flexible manner.23 As we focus most of our attention
    21
     Online Appendix 3 provides an envelope theorem result that holding z fixed yields a first-order approximation of
the compensating variation from retail price shocks under any arbitrary (unspecified) upper-tier utility function. In
the counterfactual analysis in Section 6, we report results both holding z fixed and allowing for endogenous changes
in z after specifying (1).
  22
     We show in Online Appendix 3.D that these preferences can be derived from the aggregation of discrete-choice
preferences across many individual agents within group z.
  23
     For instance, demand systems with a choke price can generate price elasticities that depend on income (Arkolakis
et al., 2018), but offer significantly less flexibility in that relationship. Related work by Feenstra & Romalis (2014)
features a single demand elasticity across incomes. Handbury (2019) derives conditions under which such preferences
are rational and well-defined, and Online Appendix 3.C provides an equivalent demand structure, based on results
in Fally (2018), in which non-homotheticities do not rely on outside good consumption z.

                                                          12
on within-product module allocations, we model the choice over product modules with a Cobb-
Douglas upper-tier, where αn (z) refers to the fraction of expenditures spent on product module n
at income level z (assuming n αn (z) = 1 for all z).24 Comparing two goods i and j within the
                            P

same module n, relative expenditures by consumers of income level z are then given by:
                                                               "                              #
                                 xni (z)                    ϕni (z)       pni
                             log         = (σn (z) − 1) log         − log                                          (3)
                                 xnj (z)                    ϕnj (z)       pnj

Motivated by the evidence discussed above, we let household quality evaluations log ϕni (z) de-
pend on an intrinsic quality term log φni associated with brand i and a multiplicative term γn (z)
depending on income level z:

                      Intrinsic Quality Assumption:                log ϕni (z) = γn (z) log φni                    (4)
                         ´
With the normalization Ωz γn (z)dz = 1 (where Ωz is a set of z household types),25 this intrinsic
quality term also corresponds to the democratic average quality evaluation across households:
                                              ˆ
                                    log φni =    log ϕni (z)dz                                (5)
                                                         Ωz

In the empirical estimation below, we estimate perceived quality ϕni (z) separately for each income
group to verify whether relative quality evaluations are indeed preserved across income levels
before imposing the above restriction. Finally, the retail price index is income-specific and given
by PG (z) = n Pn (z)αn (z) , where the price index Pn (z) for each product module n is defined as:
            Q

                                                                                1
                                                                               1−σn (z)
                                              X     1−σ (z)
                               Pn (z) =           pni n ϕni (z)σn (z)−1                 .                        (6)
                                            i∈Gn

Production For each product group n, entrepreneurs draw their productivity a from a cumula-
tive distribution Gn (a) upon paying a sunk entry cost FnE , as in Melitz (2003). For the remainder
of this section, we index firms by a instead of i, since all relevant firm-level decisions are uniquely
determined by firm productivity a. The timing of events is as follows. First, entrepreneurs pay
the entry cost FnE and discover their productivity a. Second, each entrepreneur decides at which
level of quality to produce or exit. Third, production occurs and markets clear. Firms compete in
prices, allowing for oligopolistic interactions as in Hottman et al. (2016).26
  24
      We abstract from within-brand product substitution by summing up sales across potentially multiple barcodes.
Online Appendix 4.A presents an extension of our model to multi-product firms following Hottman et al. (2016).
In our setting, we show that as long as the ratio of between to within-brand elasticities of substitution does not
significantly differ across income groups z, firm variation in within-brand product variety does not play a role for
differences in firm sizes across z (Figure 1). In Section 5.1 we find minor differences in the cross-brand elasticities
across income groups, and in Table A.5 we find no evidence of significant differences in the between-to-within ratios.
   25
      This normalization sets the simple mean of preference parameters γn (z) equal to unity across a fixed set of
household types z. As shown in Online Appendix 3.E, this normalization to unity across household types z is
without loss of generality.
   26
      We also consider a generalized version of Cournot quantity competition following Atkeson and Burstein (2008),
whereby firms compete in quantity. This yields very similar results, both for the cross-section of implied markups
as well as for the counterfactual analysis.

                                                          13
We normalize the cost of labor (wage w) to unity.27 There are two cost components: a variable
and a fixed cost (in terms of labor). We allow for the possibility that both the marginal and the
fixed cost of production are a function of output quality. The latter captures potential overhead
costs such as design, R&D and marketing which do not directly depend on the quantities being
produced but affect the quality of the product. In turn, variable costs depend on the level of
quality of the production as well as the entrepreneur’s productivity, as in Melitz (2003). Hence,
the total cost associated with the production of a quantity q with quality φ and productivity a is:

                                             cn (φ)q/a + fn (φ) + f0n                                                (7)

where fn (φ) is the part of fixed costs that directly depend on output quality. For tractability, we
adopt a simple log-linear parameterization for fixed costs as in Hallak & Sivadasan (2013) and the
second model variant of Kugler & Verhoogen (2012):
                                                                    1
                                                 fn (φ) = bn βn φ βn                                                 (8)

For instance, fixed costs increase with quality (βn > 0) if higher quality entails higher expenditures
on research and development or marketing, for which the cost is not directly dependent of the scale
of production. On the contrary, it could be the case that fixed costs are decreasing in quality if
such fixed investments are most suitable for mass-producing low-quality output at lower marginal
costs.28 Similarly, we assume that variable costs depend log-linearly on quality, with parameter
ξn ≥ 0 to capture the elasticity of the cost increase to the level of output quality:29

                                                     cn (φ) = φξn                                                    (9)

As long as ξn is smaller than the minimum quality evaluation γn (z), firms choose positive levels of
quality in equilibrium, as we further discuss below.

4.2    Equilibrium and Counterfactuals
In equilibrium, consumers maximize their utility, expected profits upon entry equal the sunk entry
cost, and firms choose their price, quality and quantity to maximize profits. There are two sources
for variable markups across firms in our setting. They are determined by both their market power
in oligopoly as well as the composition of the consumers that each firm sells to (i.e. the sales-
weighted average price elasticity they face across consumer groups z). To see this, prices are given
by:
                                                 φ(a)ξn µn (a)
                                        pn (a) =                                                (10)
                                                      a
  27
     Our focus is on the consumption side while shutting down potential implications of firm heterogeneity for nominal
earnings inequality (e.g. Helpman et al. (2017); Song et al. (2018)). In counterfactuals, GE labor market adjustments
could have additional knock-on effects on relative prices across firms and, thus, differences in inflation between income
groups, that our analysis ignores.
  28
     In Online Appendix 4.B, we also present a model extension allowing for heterogeneous fixed costs that may be
correlated with productivity, as in Hallak & Sivadasan (2013).
  29
     There is no need for a constant term as it would be isomorphic to a common productivity shifter after redefining
Gn (a).

                                                           14
where µn is the markup over marginal cost:
                                                       ´
                             pn (a)                      z xn (z, a) dH(z)
                    µn (a) ≡        =1+ ´                                                                      (11)
                             cn (a)       z (σn (z) − 1) (1 − sn (z, a)) xn (z, a) dH(z)

xn (z, a) denotes sales of firm with productivity a to consumers of income level z. sn (z, a) are
market shares in product module n. Due to oligopolistic competition, this markup is larger than
  σ̃n (a)
σ̃n (a)−1 , the markup under monopolistic competition, where σ̃n (a) is the weighted average elasticity
of substitution across consumers that the firm sells to:
                                            ´
                                              σn (z) xn (z, a)dH(z)
                                  σ̃n (a) = z ´
                                                z xn (z, a)dH(z)

The first-order condition in φ characterizes optimal quality φn (a) for firms with productivity a.
Existence of a non-degenerate equilibrium choice for quality requires that we fit in either of two
cases.30 In the first case, higher quality entails higher fixed costs (βn > 0 and bn > 0) and
the increase in marginal costs does not exceed consumer valuation for quality (ξn < γn (z) for
all z). In this case, large firms sort into producing higher quality. In the second case, marginal
costs increase more strongly with higher quality output (ξn > γn (z) for all z), and fixed costs
are decreasing (βn < 0 and bn < 0). In this case, relatively small “boutique” producers sort into
producing higher quality products while mass producers produce low-cost, low-quality products.
    In both cases, optimal quality satisfies:
                                                                         βn
                                                     γ̃n (a) − ξn
                                                 
                                      φn (a) =                    . Xn (a)                                     (12)
                                                       bn µn (a)
                 ´
where Xn (a) = z x(a, z)dH(z) denotes total sales of firm a in product module n and γ̃n (a) is the
weighted average quality valuation γn (z) for firm with productivity a, weighted by sales and price
elasticities across its customer base:
                                 ´
                                   γn (z) (σn (z)−1) (1 − sn (z, a)) xn (z, a) dH(z)
                        γ̃n (a) = z ´                                                          (13)
                                     z (σn (z)−1) (1 − sn (z, a)) xn (z, a) dH(z)

Optimal quality is determined by several forces that are apparent in equation (12). First, when
βn > 0, larger sales induce higher optimal quality, as reflected in the term Xn (a)βn . This is the
scale effect due to the higher fixed costs of producing at higher quality. If we compare two firms
with the same customer base, the larger one would more profitably invest in upgrading quality
if βn > 0. Second, optimal quality depends on how much the firm-specific customer base value
quality, captured by γ̃n (a). Firms that tend to sell to consumers with high γn (z) also tend to have
higher returns to quality upgrading. Third, optimal quality depends on technology and the cost
structure. A higher elasticity of marginal costs to quality, ξn , induces lower optimal quality. If
instead βn < 0, higher quality is associated with smaller firm scale Xn (a), a higher average quality
valuation γ̃n (a) and a lower elasticity of marginal costs to quality ξn .
  30
     If βn > 0, bn > 0 and ξn > γn (z) optimal quality is zero. If βn < 0, bn < 0 and ξn < γn (z) optimal quality is
infinite.

                                                         15
You can also read